characterize global features. Therefore, DeSTIN learns features from data in a total unsupervised fashion, which stands in contrast to the other architectures described that rely on previous knowl- edge of the problem at hand.
An enhancement to the original DeSTIN is proposed in [5], where an innovative recurrent clus- tering algorithm is employed for capturing the spatio-temporal dependencies, as an unsupervised learning procedure included in each node of the DeSTIN hierarchy. Furthermore, owing to par- allel and independent operation of each node, remarkable scalability attributes, specially in GPU implementation, are offered. Note that modern GPU are highly parallel programmable processors that have a peak arithmetic and memory bandwidth much greater than any CPU [78]. Thus, the ability to map DeSTIN to a GPU implementation is of great importance in order to tackle large problems that use high resolution datasets, such as image or video.
Capturing the temporal dependencies on data is not straightforward. Also, as mentioned ear- lier, it has not been an object of intensive research since most architectures are focused on im- proving their performance on learning spatial features. Nevertheless, the ability to extract features based on the sequential behaviour of data, in an unsupervised manner, seems very promising. In the particular field of power systems, modelling the temporal dependencies inherent to consecutive information prospects great advances on monitoring and control the dynamics of the system.
3.4
Final Remarks
The advent of WAMS and consequent PMU mass-deployment is of great importance for the improvement of monitoring and control of power systems. However, as stated in section 2.10, power system operators are craving for techniques that efficiently deal with the amount of raw data flooding control centers nowadays. Artificial Intelligence, and more specifically Deep Learn- ing approaches, seem to be a valid course of action to follow. The raw data provided by the PMUs prospects a need for unsupervised learning techniques that extract valuable information from patterns recognized in synchronized measurements. Otherwise, training supervised frame- works requires previous extensive labelling of their datasets. Due to its unsupervised pre-training phase, the Deep Belief Networks can emerge as a worthwhile framework to be employed. Alter- natively, the flexibility exhibited by Convolution Neural Networks makes them highly attractive, despite needing a way to encode data as images. Moreover, Autoencoders are envisaged as a means of data dimensionality reduction. The issue regarding missing data on PMUs might aswell be a motivation for the employment of Autoencoders, in its Denoising version. In addition, the highly temporal component of the patterns present in PMU data opens doors for the application of spatio-temporal algorithms for harnessing temporal dependencies.
Recent deployments of Deep Learning approaches regarding electric power systems in general are depicting a bright future for the coupling of both activities. Indeed, several applications have proven to benefit from the employment of deep learning frameworks. Also, the current availabil- ity of GPU implementation is assuaging the computational effort inherent to those architectures. Thus, the previous and the present section comprised an overview of both PMU technology and
Deep Learning paradigm, aiming at a revision of concepts and respective applications. The inten- tion of this dissertation is to propose an innovative application of deep learning frameworks that can effectively extract features from synchronized measurements provided by PMUs.
Chapter 4
Context of The Work
This chapter contextualizes the work developed. It describes the PMU-based measurement system that provided the data used, as well as the classification task carried out. The impacts of each disturbance in system frequency are detailed, hence providing an insight to the ongoing electrical phenomena.
4.1
The Medfasee BT Project
As described earlier in section 2.9, Brazil has implemented a Low Voltage, PMU-based, Wide Area Measurement System. By the end of 2013, 22 devices were already installed in several universities spread across the country. In addition, a 23rd PMU was envisaged to be installed but, at the time of this thesis, that still had not happened. The data collected from each PMU is sent to a central PDC installed in Universidade Federal de Santa Catarina (UFSC). So as to better understand the geographical disposal of the devices, the location of each university taking part in the project is identified in Figure4.1. The acronym of each measurement point is described in AppendixD.
Besides R&D, the data produced is being harnessed by the brazilian system operator forpost- mortemdisturbance analysis. Also, the geographical distribution of the installed devices allows the monitoring of electromechanical oscillations [6].
It is worth noticing that the Medfasee BT Project contributed greatly for the development of this thesis by providing a complete dataset with several events properly labelled. Indeed, a manual database was created including real events registered by the system operator between 2010 and 2015 [79]. Since the PMUs are installed in the low voltage grid, a careful selection of cases was performed, which highlighted four types of events:
• Generation Tripping - cases of generation loss that caused frequency changes greater than 0.08 Hz (approximately 530 MW);
• Load Shedding - cases of load loss that resulted in frequency changes greater than 0.07 Hz (approximately 440 MW);
Figure 4.1: Geographical disposal of PMUs implemented in Medfasee Project [6] - details in AppendixD
• Line Tripping - disconnection ofě500kV transmission lines with significant power flow;
• Oscillation - cases of inter-area oscillations caused by the disconnection of a specific 600kV DC link.
The events detected are distinguishable by the extent of their impacts. That is, both Generation Tripping and Load Shedding are considered systemic events, therefore reaching a larger number of PMUs. In contrast, Line Trippings and Oscillations are observed locally, hence affecting less PMUs. As a result, the number of PMUs worth accessing is highly event-dependent. Having gathered the exact time and location of each disturbance, the data measured by the affected PMUs (stored in the main PDC) was then collected using a software for offline synchrophasor analysis, which was specially developed for the project (MedPlot [33]). Table4.1lists the number of events considered for each disturbance and the respective total number of cases (which corresponds to the number of events multiplied by the number of PMUs affected in each event).
Table 4.1: List of extracted events
Events no. of events no. of examples
Generation Tripping 55 876
Load Shedding 29 421
Line Tripping 17 46
Oscillation 7 14